vm: heap-force builtin-arg wrapper to plug arena-window ref leak

[InauguralPhysicist]

Summary

• Multi-arg builtin calls inside an arena_mark/arena_reset window leaked one ref per heap argument.
• The CALL path packs args into a wrapper make_list(argc); with g_arena.active=1 that wrapper was arena-allocated. list_append incref'd heap items (e.g. weights, grad), but the matching val_decref(arg) is a no-op on arena values — arena reclaim doesn't walk children, so the items' incref was permanently stranded.
• Fix: new make_list_heap() constructor that forces heap regardless of arena state; used by the two builtin-arg wrapper sites (jit_helper_call, CASE(CALL)). val_decref(arg) now actually releases the items.

Why this matters

Minimal repro (no loop, no allocation pressure):

weights is [1.0, 2.0, 3.0, 4.0]
grad is [0.01, 0.01, 0.01, 0.01]
arena_mark of null
sgd_update of [weights, grad, 1.0]
arena_reset of null

Before: 1408 bytes leaked in 12 allocations (weights + grad lists + 10 inner items + items arrays). After: clean.

The bug generalizes: any multi-arg builtin (str of [a, b], contains of [list, x], etc.) called inside an arena window leaks one ref per heap arg position. SGD training loops accumulate one phantom ref per arg per step, so weights/grad globals never reach refcount 0 at exit.

Test plan

• make asan && cd tests && ASAN_OPTIONS=detect_leaks=1 bash run_all_tests.sh — 1882/1882, leak tally 13 (unchanged ceiling)
• make && cd tests && bash run_all_tests.sh — 1882/1882
• Standalone leak audit: 19 → 17 leakers (retires test_arena_ownership.eigs and test_optimize.eigs)
• Minimal repro verified leak-clean

🤖 Generated with Claude Code